portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 15:27:06 +01:00
Code Releases Activity

Dirichlet first cut - wrong answers on dagger multiply.

Browse Source 
Struggling to get a compute node so changing systems
This commit is contained in:
Peter Boyle
2022-02-22 19:58:33 +00:00
parent da06d15f73
commit aab3bcb46f
12 changed files with 523 additions and 440 deletions
Download Patch File Download Diff File Expand all files Collapse all files

58
Grid/communicator/Communicator_mpi3.cc
View File

@ -106,7 +106,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
// Remap using the shared memory optimising routine
// The remap creates a comm which must be freed
////////////////////////////////////////////////////
GlobalSharedMemory::OptimalCommunicator (processors,optimal_comm);
GlobalSharedMemory::OptimalCommunicator (processors,optimal_comm,_shm_processors);
InitFromMPICommunicator(processors,optimal_comm);
SetCommunicator(optimal_comm);
///////////////////////////////////////////////////
@ -124,12 +124,13 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
Coordinate parent_processor_coor(_ndimension,0);
Coordinate parent_processors (_ndimension,1);
Coordinate shm_processors (_ndimension,1);
// Can make 5d grid from 4d etc...
int pad = _ndimension-parent_ndimension;
for(int d=0;d<parent_ndimension;d++){
parent_processor_coor[pad+d]=parent._processor_coor[d];
parent_processors [pad+d]=parent._processors[d];
shm_processors [pad+d]=parent._shm_processors[d];
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
@ -154,6 +155,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
ccoor[d] = parent_processor_coor[d] % processors[d];
scoor[d] = parent_processor_coor[d] / processors[d];
ssize[d] = parent_processors[d] / processors[d];
if ( processors[d] < shm_processors[d] ) shm_processors[d] = processors[d]; // subnode splitting.
}
// rank within subcomm ; srank is rank of subcomm within blocks of subcomms
@ -335,22 +337,22 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
}
// Basic Halo comms primitive
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int dest,
int dest, int dox,
void *recv,
int from,
int from, int dor,
int bytes,int dir)
{
std::vector<CommsRequest_t> list;
double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,dir);
StencilSendToRecvFromComplete(list,dir);
return offbytes;
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
int dest,int dox,
void *recv,
int from,
int from,int dor,
int bytes,int dir)
{
int ncomm =communicator_halo.size();
@ -370,28 +372,32 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
double off_node_bytes=0.0;
int tag;
if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+from*32;
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
assert(ierr==0);
list.push_back(rrq);
off_node_bytes+=bytes;
if ( dox ) {
if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+from*32;
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
assert(ierr==0);
list.push_back(rrq);
off_node_bytes+=bytes;
}
}
if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+_processor*32;
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
assert(ierr==0);
list.push_back(xrq);
off_node_bytes+=bytes;
} else {
if (dor) {
if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+_processor*32;
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
assert(ierr==0);
list.push_back(xrq);
off_node_bytes+=bytes;
} else {
// TODO : make a OMP loop on CPU, call threaded bcopy
void *shm = (void *) this->ShmBufferTranslate(dest,recv);
assert(shm!=NULL);
// std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
void *shm = (void *) this->ShmBufferTranslate(dest,recv);
assert(shm!=NULL);
// std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
}
}
if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
this->StencilSendToRecvFromComplete(list,dir);
}